I'm not sure why anyone would even do that in the first place, LLama doesn't generate synthetic data that would be even remotely good enough. Even GPT 3.5 and 4 are already very borderline for it, with lots of wrong and censored answers. And at best you make a model that's as good as LLama is, i.e. not very.
Instruction-tuning is the obvious use case. That much has nothing to do with subjectivity, alignment or censorship, it's will-you-actually-show-this-as-JSON-if-asked.
That's tuning llama which is allowed from what I understand.
Otherwise why release it at all, it's not very functional in its initial state anyway. What that applies to is using llama outputs to train a completely new base model which makes no practical sense.
As for generating jsons, that's more of a inference runtime thing, since you need to pick the top tokens that result a valid json instead of just hoping it returns something that can be parsed. On top of extensive tuning of course.
